A Real-Time Car Towing Management System Using ML-Powered Automatic Number Plate Recognition

Round 1

Reviewer 1 Report

Comments are attached in pdf file.

Comments for author File: Comments.pdf

Author Response

Reviewer#1, Concern # 1: This complete process can be computed using fully convolutional layer network so explain why using multiple stages having conventional image processing approach and mathematical techniques/approaches to detect number plates and characters separately is more appropriate.

Author response: Thanks for the comment. We selected KNN because it is a lightweight algorithm that can be deployed on mobile devices, which can outperform the state-of-the-art ANPR approaches in terms of the overall processing time. Training and embedding deep learning models (e.g., CNN, RNN, ANN, etc.) into a mobile app can be challenging because of the limited resources of handheld devices. Furthermore, these DL models require large and robust datasets of plate photos of different sizes, orientations, and shapes across different regions worldwide to train them. Moreover, these datasets may contain sensitive identifying information related to the vehicle, driver, and location, making it a challenging task to acquire

Reviewer#1, Concern # 2: If the dataset is limited, data augmentation or synthetic dataset can be used to train a convolutional neural network that will provide better classification accuracy. Reason the computational overhead of augmentation and synthetic process to justify the proposed system.

Author response: Thanks for the comment. For the character recognition phase, we needed to train only the alphanumeric characters, which be easily learned from our dataset. We didn’t find it is necessary to perform any dataset augmentation because of its computational overhead. Again, one of our research objectives is to use lightweight algorithms that run on mobile devices with high performance.

Reviewer#1, Concern # 3: It is mentioned that network is trained for different lighting conditions, some figures/results with low lighting conditions should also be added.

Author response: Thanks for the comment.

Reviewer#1, Concern # 4: The architecture proposed in this paper needs to be compared with state-of-the-art networks (like residual networks) that are faster to train, more accurate and have fewer parameters. Even if it is not possible to compare complete architecture, at least compare KNN based character recognition system with MLP based Character recognition system as that network can be trained with dataset like MNIST for recognizing character and can be more accurate than KNN based system.

Author response Thanks for the comment. Training ANN models (e.g., residual networks) would be resource-intensive, which can degrade the mobile app's performance. Therefore, we compared the performance of our ANPR system to some existing ML models (i.e., SVM) and computer vision models (i.e., YOLO detector). We found that our ANPR system outperforms SVM and YOLO in terms of processing time, which is critical for the ANPR applications which need to operate in real-time.

Reviewer#1, Concern # 5: The similarity index of the paper is 40 percent. Even if it is from your own old work, try to reduce the similar components.

Author response: Thanks for the comment. We reduced the similarity index significantly throughout the whole paper to address this comment.

Reviewer 2 Report

This paper proposes a mobile-based Automatic Number Plate Recognition (ANPR) system to enhance the accuracy of recognizing number plates.
Here are some comments which can help to improve the paper.

1) The main research problem in this paper is the accuracy of the ANPR as mentioned in the Abstract. However, the Introduction says the main problem is the privacy of the large datasets needed in the methods provided in the literature. I am a little bit confused about the research problem. The authors say they used KNN in the second phase of the proposed method and we know that the accuracy of KNN depends on the provided datasets. Now, the question of the privacy of the dataset still exists. I suggest authors clearly discuss this in both Abstract and Introduction.

2) For Section 3, I suggest authors to provided a high-level architecture of the proposed method including the main phases and the main modules in each phase.

3) In Section 5, the authors provided the evaluation results based on the datasets they collected from two sources. I am curious to see the evaluation results against the datasets used in the original papers [13] and [16]. At least the authors must provide the evaluation results based on some standard and public datasets to show fair comparison.

Author Response

Reviewer#2, Concern # 1: The main research problem in this paper is the accuracy of the ANPR as mentioned in the Abstract. However, the Introduction says the main problem is the privacy of the large datasets needed in the methods provided in the literature. I am a little bit confused about the research problem. The authors say they used KNN in the second phase of the proposed method and we know that the accuracy of KNN depends on the provided datasets. Now, the question of the privacy of the dataset still exists. I suggest authors clearly discuss this in both Abstract and Introduction.

Author response: Thanks for the comment. We addressed this concern in the abstract and introduction sections. The main research problem of this work is to provide an alternative ANPR approach that can accurately identify the plate’s area and recognize its alphanumeric characters in real-time without the need for substantial ANPR datasets that may contain sensitive information related to car owners to train ML models. Our ANPR system has two main modules for detecting the number plate’s location on the car body and identifying the plate’s alphanumeric characters, respectively. First, we developed an algorithm for sketching a bounding box around the input image’s plate. Second, the plate box is then fed to the second module, which generates the contours for all alphanumeric characters within the plate box. We used KNN to recognize the actual characters and digits in the number plate. We selected KNN because it is a lightweight algorithm that can be deployed on mobile devices, which can outperform the state-of-the-art ANPR approaches in terms of the overall processing time.

 

Reviewer#2, Concern # 2: For Section 3, I suggest authors to provide a high-level architecture of the proposed method including the main phases and the main modules in each phase.

Author response: Thanks for the comment. We added Figure 3 that illustrates the system architecture of the car towing management system. Also, Figure 2 and Algorithms 1 & 2 describe the phases of the plate detection, character segmentation, and recognition modules.

 

Reviewer#2, Concern # 1: In Section 5, the authors provided the evaluation results based on the datasets they collected from two sources. I am curious to see the evaluation results against the datasets used in the original papers [13] and [16]. At least the authors must provide the evaluation results based on some standard and public datasets to show fair comparison.

Author response: Thanks for the comment. We tested our model using a fresh image collected from different sources such as Kaggle and Google Web Scraper. These images haven’t been used in the training phase. Many images in our dataset are in their natural environments because plate detection is highly dependent on contextual information. The main reason for comparing our model to SVM [16] and YOLO detector [13] is to evaluate the performance of our mobile app in terms of processing time, not to compare the classification accuracy.

Round 2

Reviewer 1 Report

Overall, all my comments and suggestions have been addressed. Thanks

Author Response

Thank you for your comments.

Reviewer 2 Report

The authors answered my comments provided in the previous round of reviews.

Author Response

Thank you for your comments.